CATArena Advances AI Agent Testing While Denario Simplifies Financial Research

Researchers have made significant advancements in various areas of artificial intelligence, including executable world models, agent routing, and logical rule induction. A study on executable world models for ARC-AGI-3 agents achieved 90% relative human action efficiency on 6 games and fully solved 7 games. Another study introduced Uno-Orchestra, a unified orchestration policy that selectively decomposes tasks and dispatches subtasks to admissible model-primitive pairs, reaching 77% macro pass@1 on a 13-benchmark suite. A foundation model for zero-shot logical rule induction, Neural Rule Inducer (NRI), was introduced, which represents literals using domain-agnostic statistical properties and achieves competitive or superior predictive performance on classical ILP benchmarks and large-scale knowledge bases.

The scaling properties of implicit deductive reasoning in Transformers were investigated, and a novel neuro-symbolic question-answering framework governed by a Probabilistic Inconsistency Signal (PIS) was introduced, which strictly decouples semantic extraction from a symbolic reasoning engine and achieves perfect 1.0 accuracy on temporal arithmetic benchmarks. A study on long-horizon search agents proposed an elastic context orchestration paradigm, Context-ReAct, which integrates reasoning, context management, and tool use in a unified loop and achieves 61.5% on BrowseComp and 62.5% on BrowseComp-ZH. A multi-agent collaborative framework for robust IMU activity recognition, SensingAgents, was introduced, which organizes LLM-powered agents into specialized roles and achieves an accuracy of 79.5% in a zero setting.

Researchers also explored the challenges of embodied AI, including the need for a privacy-utility trade-off, and proposed a unified privacy-aware framework, Secure Privacy Integration in Next-generation Embodied AI (SPINE), which treats privacy as a dynamic control signal governing cross-stage coupling throughout the entire EAI life cycle. A study on on-line learning in Tree MDPs proposed a novel approach that treats policies as bandit arms and achieves instance-dependent upper bounds on sample complexity and regret. Finally, a study on regularized centered emphatic temporal difference learning proposed a method that preserves the follow-on trace and regularizes only the auxiliary centering recursion, corresponding to lifting the lower-right block of the coupled key matrix from 1 to 1+c.